University of Twente Student Theses


Modular Neural Networks using Multiple Gradients

Engelenhoven, Adjorn van (2020) Modular Neural Networks using Multiple Gradients.

[img] PDF
Abstract:Artificial neural networks are often not able to show how they generate their results. By defining subtasks the interpretability of neural networks can be improved. Modular neural networks can use hyperparameters to define subtasks with which it should solve the problems it is trained for. However, parts of the Modular Neural Network are trained solely on the accuracy of their subtask and not the amount of useful information this subtask can provide for the entire task. In this paper the Shared Layer Modular Neural Network is proposed which could be an improvement on the standard Modular Neural Network. The accuracy could be improved by combining the layers before the output of the subtask so more relevant information of these subtasks can be combined. Furthermore, the accuracy could be improved by optimizing the weights based on a combination of the loss functions of both the subtask and the global task. In this paper, the Shared Layer Modular Neural Network is developed and tested to see if it is an improvement compared to other models. CIFAR-10 was used in the evaluation of the model. If this model is an improvement, a more accurate explainable neural network has been created which can help solve problems that are currently not completely understood.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science BSc (56964)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page